X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FPowerPC%2FPPCISelDAGToDAG.cpp;h=e3814f728af80a0b103a716bc99d1c968aeeac39;hb=cf0db29df20d9c665da7e82bb261bdd7cf7f1b2b;hp=7d7f557e433022748f61e074934efdccb85324e7;hpb=9cad6c8a24ea0f87953815c6005ddd52789c7dde;p=oota-llvm.git diff --git a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp index 7d7f557e433..e3814f728af 100644 --- a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp +++ b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp @@ -42,11 +42,15 @@ using namespace llvm; cl::opt ANDIGlueBug("expose-ppc-andi-glue-bug", cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden); -cl::opt UseBitPermRewriter("ppc-use-bit-perm-rewriter", cl::init(true), - cl::desc("use aggressive ppc isel for bit permutations"), cl::Hidden); -cl::opt BPermRewriterNoMasking("ppc-bit-perm-rewriter-stress-rotates", - cl::desc("stress rotate selection in aggressive ppc isel for " - "bit permutations"), cl::Hidden); +static cl::opt + UseBitPermRewriter("ppc-use-bit-perm-rewriter", cl::init(true), + cl::desc("use aggressive ppc isel for bit permutations"), + cl::Hidden); +static cl::opt BPermRewriterNoMasking( + "ppc-bit-perm-rewriter-stress-rotates", + cl::desc("stress rotate selection in aggressive ppc isel for " + "bit permutations"), + cl::Hidden); namespace llvm { void initializePPCDAGToDAGISelPass(PassRegistry&); @@ -59,22 +63,20 @@ namespace { /// class PPCDAGToDAGISel : public SelectionDAGISel { const PPCTargetMachine &TM; - const PPCTargetLowering *PPCLowering; const PPCSubtarget *PPCSubTarget; + const PPCTargetLowering *PPCLowering; unsigned GlobalBaseReg; public: explicit PPCDAGToDAGISel(PPCTargetMachine &tm) - : SelectionDAGISel(tm), TM(tm), - PPCLowering(TM.getSubtargetImpl()->getTargetLowering()), - PPCSubTarget(TM.getSubtargetImpl()) { + : SelectionDAGISel(tm), TM(tm) { initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry()); } bool runOnMachineFunction(MachineFunction &MF) override { // Make sure we re-emit a set of the global base reg if necessary GlobalBaseReg = 0; - PPCLowering = TM.getSubtargetImpl()->getTargetLowering(); - PPCSubTarget = TM.getSubtargetImpl(); + PPCSubTarget = &MF.getSubtarget(); + PPCLowering = PPCSubTarget->getTargetLowering(); SelectionDAGISel::runOnMachineFunction(MF); if (!PPCSubTarget->isSVR4ABI()) @@ -88,28 +90,21 @@ namespace { /// getI32Imm - Return a target constant with the specified value, of type /// i32. - inline SDValue getI32Imm(unsigned Imm) { - return CurDAG->getTargetConstant(Imm, MVT::i32); + inline SDValue getI32Imm(unsigned Imm, SDLoc dl) { + return CurDAG->getTargetConstant(Imm, dl, MVT::i32); } /// getI64Imm - Return a target constant with the specified value, of type /// i64. - inline SDValue getI64Imm(uint64_t Imm) { - return CurDAG->getTargetConstant(Imm, MVT::i64); + inline SDValue getI64Imm(uint64_t Imm, SDLoc dl) { + return CurDAG->getTargetConstant(Imm, dl, MVT::i64); } /// getSmallIPtrImm - Return a target constant of pointer type. - inline SDValue getSmallIPtrImm(unsigned Imm) { - return CurDAG->getTargetConstant(Imm, PPCLowering->getPointerTy()); + inline SDValue getSmallIPtrImm(unsigned Imm, SDLoc dl) { + return CurDAG->getTargetConstant(Imm, dl, PPCLowering->getPointerTy()); } - /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s - /// with any number of 0s on either side. The 1s are allowed to wrap from - /// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. - /// 0x0F0F0000 is not, since all 1s are not contiguous. - static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME); - - /// isRotateAndMask - Returns true if Mask and Shift can be folded into a /// rotate and mask opcode and mask operation. static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask, @@ -184,20 +179,35 @@ namespace { /// register can be improved, but it is wrong to substitute Reg+Reg for /// Reg in an asm, because the load or store opcode would have to change. bool SelectInlineAsmMemoryOperand(const SDValue &Op, - char ConstraintCode, + unsigned ConstraintID, std::vector &OutOps) override { - // We need to make sure that this one operand does not end up in r0 - // (because we might end up lowering this as 0(%op)). - const TargetRegisterInfo *TRI = TM.getSubtargetImpl()->getRegisterInfo(); - const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1); - SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32); - SDValue NewOp = - SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, - SDLoc(Op), Op.getValueType(), - Op, RC), 0); - - OutOps.push_back(NewOp); - return false; + + switch(ConstraintID) { + default: + errs() << "ConstraintID: " << ConstraintID << "\n"; + llvm_unreachable("Unexpected asm memory constraint"); + case InlineAsm::Constraint_es: + case InlineAsm::Constraint_i: + case InlineAsm::Constraint_m: + case InlineAsm::Constraint_o: + case InlineAsm::Constraint_Q: + case InlineAsm::Constraint_Z: + case InlineAsm::Constraint_Zy: + // We need to make sure that this one operand does not end up in r0 + // (because we might end up lowering this as 0(%op)). + const TargetRegisterInfo *TRI = PPCSubTarget->getRegisterInfo(); + const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1); + SDLoc dl(Op); + SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32); + SDValue NewOp = + SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS, + dl, Op.getValueType(), + Op, RC), 0); + + OutOps.push_back(NewOp); + return false; + } + return true; } void InsertVRSaveCode(MachineFunction &MF); @@ -217,11 +227,14 @@ private: void PeepholeCROps(); SDValue combineToCMPB(SDNode *N); + void foldBoolExts(SDValue &Res, SDNode *&N); bool AllUsersSelectZero(SDNode *N); void SwapAllSelectUsers(SDNode *N); + + SDNode *transferMemOperands(SDNode *N, SDNode *Result); }; -} +} // namespace /// InsertVRSaveCode - Once the entire function has been instruction selected, /// all virtual registers are created and all machine instructions are built, @@ -257,7 +270,7 @@ void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) { unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); - const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo(); + const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo(); MachineBasicBlock &EntryBB = *Fn.begin(); DebugLoc dl; // Emit the following code into the entry block: @@ -293,7 +306,7 @@ void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) { /// SDNode *PPCDAGToDAGISel::getGlobalBaseReg() { if (!GlobalBaseReg) { - const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo(); + const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo(); // Insert the set of GlobalBaseReg into the first MBB of the function MachineBasicBlock &FirstMBB = MF->front(); MachineBasicBlock::iterator MBBI = FirstMBB.begin(); @@ -306,12 +319,13 @@ SDNode *PPCDAGToDAGISel::getGlobalBaseReg() { if (M->getPICLevel() == PICLevel::Small) { BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MoveGOTtoLR)); BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); + MF->getInfo()->setUsesPICBase(true); } else { BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR)); BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg); unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass); BuildMI(FirstMBB, MBBI, dl, - TII.get(PPC::UpdateGBR)).addReg(GlobalBaseReg) + TII.get(PPC::UpdateGBR), GlobalBaseReg) .addReg(TempReg, RegState::Define).addReg(GlobalBaseReg); MF->getInfo()->setUsesPICBase(true); } @@ -393,33 +407,9 @@ SDNode *PPCDAGToDAGISel::getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset) { unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8; if (SN->hasOneUse()) return CurDAG->SelectNodeTo(SN, Opc, N->getValueType(0), TFI, - getSmallIPtrImm(Offset)); + getSmallIPtrImm(Offset, dl)); return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI, - getSmallIPtrImm(Offset)); -} - -bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) { - if (!Val) - return false; - - if (isShiftedMask_32(Val)) { - // look for the first non-zero bit - MB = countLeadingZeros(Val); - // look for the first zero bit after the run of ones - ME = countLeadingZeros((Val - 1) ^ Val); - return true; - } else { - Val = ~Val; // invert mask - if (isShiftedMask_32(Val)) { - // effectively look for the first zero bit - ME = countLeadingZeros(Val) - 1; - // effectively look for the first one bit after the run of zeros - MB = countLeadingZeros((Val - 1) ^ Val) + 1; - return true; - } - } - // no run present - return false; + getSmallIPtrImm(Offset, dl)); } bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask, @@ -534,8 +524,8 @@ SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) { } SH &= 31; - SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB), - getI32Imm(ME) }; + SDValue Ops[] = { Op0, Op1, getI32Imm(SH, dl), getI32Imm(MB, dl), + getI32Imm(ME, dl) }; return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); } } @@ -663,8 +653,8 @@ static SDNode *SelectInt64Direct(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) { unsigned Lo = Imm & 0xFFFF; unsigned Hi = (Imm >> 16) & 0xFFFF; - auto getI32Imm = [CurDAG](unsigned Imm) { - return CurDAG->getTargetConstant(Imm, MVT::i32); + auto getI32Imm = [CurDAG, dl](unsigned Imm) { + return CurDAG->getTargetConstant(Imm, dl, MVT::i32); }; // Simple value. @@ -754,8 +744,8 @@ static SDNode *SelectInt64(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) { if (!RMin) return SelectInt64Direct(CurDAG, dl, Imm); - auto getI32Imm = [CurDAG](unsigned Imm) { - return CurDAG->getTargetConstant(Imm, MVT::i32); + auto getI32Imm = [CurDAG, dl](unsigned Imm) { + return CurDAG->getTargetConstant(Imm, dl, MVT::i32); }; SDValue Val = SDValue(SelectInt64Direct(CurDAG, dl, MatImm), 0); @@ -1205,8 +1195,8 @@ class BitPermutationSelector { } } - SDValue getI32Imm(unsigned Imm) { - return CurDAG->getTargetConstant(Imm, MVT::i32); + SDValue getI32Imm(unsigned Imm, SDLoc dl) { + return CurDAG->getTargetConstant(Imm, dl, MVT::i32); } uint64_t getZerosMask() { @@ -1278,7 +1268,8 @@ class BitPermutationSelector { SDValue VRot; if (VRI.RLAmt) { SDValue Ops[] = - { VRI.V, getI32Imm(VRI.RLAmt), getI32Imm(0), getI32Imm(31) }; + { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl), + getI32Imm(31, dl) }; VRot = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); } else { @@ -1288,10 +1279,10 @@ class BitPermutationSelector { SDValue ANDIVal, ANDISVal; if (ANDIMask != 0) ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32, - VRot, getI32Imm(ANDIMask)), 0); + VRot, getI32Imm(ANDIMask, dl)), 0); if (ANDISMask != 0) ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32, - VRot, getI32Imm(ANDISMask)), 0); + VRot, getI32Imm(ANDISMask, dl)), 0); SDValue TotalVal; if (!ANDIVal) @@ -1337,8 +1328,10 @@ class BitPermutationSelector { if (VRI.RLAmt) { if (InstCnt) *InstCnt += 1; SDValue Ops[] = - { VRI.V, getI32Imm(VRI.RLAmt), getI32Imm(0), getI32Imm(31) }; - Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); + { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl), + getI32Imm(31, dl) }; + Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), + 0); } else { Res = VRI.V; } @@ -1358,13 +1351,15 @@ class BitPermutationSelector { for (auto &BG : BitGroups) { if (!Res) { SDValue Ops[] = - { BG.V, getI32Imm(BG.RLAmt), getI32Imm(Bits.size() - BG.EndIdx - 1), - getI32Imm(Bits.size() - BG.StartIdx - 1) }; + { BG.V, getI32Imm(BG.RLAmt, dl), + getI32Imm(Bits.size() - BG.EndIdx - 1, dl), + getI32Imm(Bits.size() - BG.StartIdx - 1, dl) }; Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); } else { SDValue Ops[] = - { Res, BG.V, getI32Imm(BG.RLAmt), getI32Imm(Bits.size() - BG.EndIdx - 1), - getI32Imm(Bits.size() - BG.StartIdx - 1) }; + { Res, BG.V, getI32Imm(BG.RLAmt, dl), + getI32Imm(Bits.size() - BG.EndIdx - 1, dl), + getI32Imm(Bits.size() - BG.StartIdx - 1, dl) }; Res = SDValue(CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops), 0); } } @@ -1383,10 +1378,10 @@ class BitPermutationSelector { SDValue ANDIVal, ANDISVal; if (ANDIMask != 0) ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32, - Res, getI32Imm(ANDIMask)), 0); + Res, getI32Imm(ANDIMask, dl)), 0); if (ANDISMask != 0) ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32, - Res, getI32Imm(ANDISMask)), 0); + Res, getI32Imm(ANDISMask, dl)), 0); if (!ANDIVal) Res = ANDISVal; @@ -1437,27 +1432,27 @@ class BitPermutationSelector { assert(InstMaskStart >= 32 && "Mask cannot start out of range"); assert(InstMaskEnd >= 32 && "Mask cannot end out of range"); SDValue Ops[] = - { V, getI32Imm(RLAmt), getI32Imm(InstMaskStart - 32), - getI32Imm(InstMaskEnd - 32) }; + { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl), + getI32Imm(InstMaskEnd - 32, dl) }; return SDValue(CurDAG->getMachineNode(PPC::RLWINM8, dl, MVT::i64, Ops), 0); } if (InstMaskEnd == 63) { SDValue Ops[] = - { V, getI32Imm(RLAmt), getI32Imm(InstMaskStart) }; + { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) }; return SDValue(CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, Ops), 0); } if (InstMaskStart == 0) { SDValue Ops[] = - { V, getI32Imm(RLAmt), getI32Imm(InstMaskEnd) }; + { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskEnd, dl) }; return SDValue(CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Ops), 0); } if (InstMaskEnd == 63 - RLAmt) { SDValue Ops[] = - { V, getI32Imm(RLAmt), getI32Imm(InstMaskStart) }; + { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) }; return SDValue(CurDAG->getMachineNode(PPC::RLDIC, dl, MVT::i64, Ops), 0); } @@ -1498,15 +1493,15 @@ class BitPermutationSelector { assert(InstMaskStart >= 32 && "Mask cannot start out of range"); assert(InstMaskEnd >= 32 && "Mask cannot end out of range"); SDValue Ops[] = - { Base, V, getI32Imm(RLAmt), getI32Imm(InstMaskStart - 32), - getI32Imm(InstMaskEnd - 32) }; + { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl), + getI32Imm(InstMaskEnd - 32, dl) }; return SDValue(CurDAG->getMachineNode(PPC::RLWIMI8, dl, MVT::i64, Ops), 0); } if (InstMaskEnd == 63 - RLAmt) { SDValue Ops[] = - { Base, V, getI32Imm(RLAmt), getI32Imm(InstMaskStart) }; + { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) }; return SDValue(CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops), 0); } @@ -1653,10 +1648,10 @@ class BitPermutationSelector { SDValue ANDIVal, ANDISVal; if (ANDIMask != 0) ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64, - VRot, getI32Imm(ANDIMask)), 0); + VRot, getI32Imm(ANDIMask, dl)), 0); if (ANDISMask != 0) ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64, - VRot, getI32Imm(ANDISMask)), 0); + VRot, getI32Imm(ANDISMask, dl)), 0); if (!ANDIVal) TotalVal = ANDISVal; @@ -1803,10 +1798,10 @@ class BitPermutationSelector { SDValue ANDIVal, ANDISVal; if (ANDIMask != 0) ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64, - Res, getI32Imm(ANDIMask)), 0); + Res, getI32Imm(ANDIMask, dl)), 0); if (ANDISMask != 0) ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64, - Res, getI32Imm(ANDISMask)), 0); + Res, getI32Imm(ANDISMask, dl)), 0); if (!ANDIVal) Res = ANDISVal; @@ -1951,11 +1946,13 @@ SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, // SETEQ/SETNE comparison with 16-bit immediate, fold it. if (isUInt<16>(Imm)) return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, - getI32Imm(Imm & 0xFFFF)), 0); + getI32Imm(Imm & 0xFFFF, dl)), + 0); // If this is a 16-bit signed immediate, fold it. if (isInt<16>((int)Imm)) return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, - getI32Imm(Imm & 0xFFFF)), 0); + getI32Imm(Imm & 0xFFFF, dl)), + 0); // For non-equality comparisons, the default code would materialize the // constant, then compare against it, like this: @@ -1967,21 +1964,22 @@ SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, // cmplwi cr0,r0,0x5678 // beq cr0,L6 SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS, - getI32Imm(Imm >> 16)), 0); + getI32Imm(Imm >> 16, dl)), 0); return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor, - getI32Imm(Imm & 0xFFFF)), 0); + getI32Imm(Imm & 0xFFFF, dl)), 0); } Opc = PPC::CMPLW; } else if (ISD::isUnsignedIntSetCC(CC)) { if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm)) return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS, - getI32Imm(Imm & 0xFFFF)), 0); + getI32Imm(Imm & 0xFFFF, dl)), 0); Opc = PPC::CMPLW; } else { short SImm; if (isIntS16Immediate(RHS, SImm)) return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS, - getI32Imm((int)SImm & 0xFFFF)), + getI32Imm((int)SImm & 0xFFFF, + dl)), 0); Opc = PPC::CMPW; } @@ -1992,11 +1990,13 @@ SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, // SETEQ/SETNE comparison with 16-bit immediate, fold it. if (isUInt<16>(Imm)) return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, - getI32Imm(Imm & 0xFFFF)), 0); + getI32Imm(Imm & 0xFFFF, dl)), + 0); // If this is a 16-bit signed immediate, fold it. if (isInt<16>(Imm)) return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, - getI32Imm(Imm & 0xFFFF)), 0); + getI32Imm(Imm & 0xFFFF, dl)), + 0); // For non-equality comparisons, the default code would materialize the // constant, then compare against it, like this: @@ -2009,22 +2009,23 @@ SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, // beq cr0,L6 if (isUInt<32>(Imm)) { SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS, - getI64Imm(Imm >> 16)), 0); + getI64Imm(Imm >> 16, dl)), 0); return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor, - getI64Imm(Imm & 0xFFFF)), 0); + getI64Imm(Imm & 0xFFFF, dl)), + 0); } } Opc = PPC::CMPLD; } else if (ISD::isUnsignedIntSetCC(CC)) { if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm)) return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS, - getI64Imm(Imm & 0xFFFF)), 0); + getI64Imm(Imm & 0xFFFF, dl)), 0); Opc = PPC::CMPLD; } else { short SImm; if (isIntS16Immediate(RHS, SImm)) return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS, - getI64Imm(SImm & 0xFFFF)), + getI64Imm(SImm & 0xFFFF, dl)), 0); Opc = PPC::CMPD; } @@ -2099,7 +2100,7 @@ static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) { // getVCmpInst: return the vector compare instruction for the specified // vector type and condition code. Since this is for altivec specific code, -// only support the altivec types (v16i8, v8i16, v4i32, and v4f32). +// only support the altivec types (v16i8, v8i16, v4i32, v2i64, and v4f32). static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC, bool HasVSX, bool &Swap, bool &Negate) { Swap = false; @@ -2178,6 +2179,8 @@ static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC, return PPC::VCMPEQUH; else if (VecVT == MVT::v4i32) return PPC::VCMPEQUW; + else if (VecVT == MVT::v2i64) + return PPC::VCMPEQUD; break; case ISD::SETGT: if (VecVT == MVT::v16i8) @@ -2186,6 +2189,8 @@ static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC, return PPC::VCMPGTSH; else if (VecVT == MVT::v4i32) return PPC::VCMPGTSW; + else if (VecVT == MVT::v2i64) + return PPC::VCMPGTSD; break; case ISD::SETUGT: if (VecVT == MVT::v16i8) @@ -2194,6 +2199,8 @@ static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC, return PPC::VCMPGTUH; else if (VecVT == MVT::v4i32) return PPC::VCMPGTUW; + else if (VecVT == MVT::v2i64) + return PPC::VCMPGTUD; break; default: break; @@ -2220,26 +2227,29 @@ SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) { default: break; case ISD::SETEQ: { Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0); - SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) }; + SDValue Ops[] = { Op, getI32Imm(27, dl), getI32Imm(5, dl), + getI32Imm(31, dl) }; return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); } case ISD::SETNE: { if (isPPC64) break; SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, - Op, getI32Imm(~0U)), 0); + Op, getI32Imm(~0U, dl)), 0); return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op, AD.getValue(1)); } case ISD::SETLT: { - SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; + SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl), + getI32Imm(31, dl) }; return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); } case ISD::SETGT: { SDValue T = SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0); T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0); - SDValue Ops[] = { T, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; + SDValue Ops[] = { T, getI32Imm(1, dl), getI32Imm(31, dl), + getI32Imm(31, dl) }; return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); } } @@ -2250,34 +2260,35 @@ SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) { case ISD::SETEQ: if (isPPC64) break; Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, - Op, getI32Imm(1)), 0); + Op, getI32Imm(1, dl)), 0); return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32, SDValue(CurDAG->getMachineNode(PPC::LI, dl, MVT::i32, - getI32Imm(0)), 0), - Op.getValue(1)); + getI32Imm(0, dl)), + 0), Op.getValue(1)); case ISD::SETNE: { if (isPPC64) break; Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0); SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, - Op, getI32Imm(~0U)); + Op, getI32Imm(~0U, dl)); return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0), Op, SDValue(AD, 1)); } case ISD::SETLT: { SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op, - getI32Imm(1)), 0); + getI32Imm(1, dl)), 0); SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD, Op), 0); - SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; + SDValue Ops[] = { AN, getI32Imm(1, dl), getI32Imm(31, dl), + getI32Imm(31, dl) }; return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); } case ISD::SETGT: { - SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) }; - Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), - 0); + SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl), + getI32Imm(31, dl) }; + Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op, - getI32Imm(1)); + getI32Imm(1, dl)); } } } @@ -2289,6 +2300,9 @@ SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) { // Altivec Vector compare instructions do not set any CR register by default and // vector compare operations return the same type as the operands. if (LHS.getValueType().isVector()) { + if (PPCSubTarget->hasQPX()) + return nullptr; + EVT VecVT = LHS.getValueType(); bool Swap, Negate; unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC, @@ -2324,15 +2338,23 @@ SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) { IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg, CCReg), 0); - SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31), - getI32Imm(31), getI32Imm(31) }; + SDValue Ops[] = { IntCR, getI32Imm((32 - (3 - Idx)) & 31, dl), + getI32Imm(31, dl), getI32Imm(31, dl) }; if (!Inv) return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); // Get the specified bit. SDValue Tmp = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0); - return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1)); + return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1, dl)); +} + +SDNode *PPCDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) { + // Transfer memoperands. + MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1); + MemOp[0] = cast(N)->getMemOperand(); + cast(Result)->setMemRefs(MemOp, MemOp + 1); + return Result; } @@ -2392,7 +2414,8 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue ShiftAmt = CurDAG->getTargetConstant(*cast(N->getOperand(1))-> - getConstantIntValue(), N->getValueType(0)); + getConstantIntValue(), dl, + N->getValueType(0)); if (N->getValueType(0) == MVT::i64) { SDNode *Op = CurDAG->getMachineNode(PPC::SRADI, dl, MVT::i64, MVT::Glue, @@ -2453,9 +2476,10 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { SDValue Chain = LD->getChain(); SDValue Base = LD->getBasePtr(); SDValue Ops[] = { Offset, Base, Chain }; - return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0), - PPCLowering->getPointerTy(), - MVT::Other, Ops); + return transferMemOperands(N, CurDAG->getMachineNode(Opcode, dl, + LD->getValueType(0), + PPCLowering->getPointerTy(), + MVT::Other, Ops)); } else { unsigned Opcode; bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; @@ -2464,6 +2488,8 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load"); switch (LoadedVT.getSimpleVT().SimpleTy) { default: llvm_unreachable("Invalid PPC load type!"); + case MVT::v4f64: Opcode = PPC::QVLFDUX; break; // QPX + case MVT::v4f32: Opcode = PPC::QVLFSUX; break; // QPX case MVT::f64: Opcode = PPC::LFDUX; break; case MVT::f32: Opcode = PPC::LFSUX; break; case MVT::i32: Opcode = PPC::LWZUX; break; @@ -2488,9 +2514,10 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { SDValue Chain = LD->getChain(); SDValue Base = LD->getBasePtr(); SDValue Ops[] = { Base, Offset, Chain }; - return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0), - PPCLowering->getPointerTy(), - MVT::Other, Ops); + return transferMemOperands(N, CurDAG->getMachineNode(Opcode, dl, + LD->getValueType(0), + PPCLowering->getPointerTy(), + MVT::Other, Ops)); } } @@ -2503,7 +2530,8 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { if (isInt32Immediate(N->getOperand(1), Imm) && isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) { SDValue Val = N->getOperand(0).getOperand(0); - SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; + SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl), + getI32Imm(ME, dl) }; return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); } // If this is just a masked value where the input is not handled above, and @@ -2512,14 +2540,15 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { isRunOfOnes(Imm, MB, ME) && N->getOperand(0).getOpcode() != ISD::ROTL) { SDValue Val = N->getOperand(0); - SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) }; + SDValue Ops[] = { Val, getI32Imm(0, dl), getI32Imm(MB, dl), + getI32Imm(ME, dl) }; return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); } // If this is a 64-bit zero-extension mask, emit rldicl. if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) && isMask_64(Imm64)) { SDValue Val = N->getOperand(0); - MB = 64 - CountTrailingOnes_64(Imm64); + MB = 64 - countTrailingOnes(Imm64); SH = 0; // If the operand is a logical right shift, we can fold it into this @@ -2534,7 +2563,7 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { SH = 64 - Imm; } - SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB) }; + SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl) }; return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops); } // AND X, 0 -> 0, not "rlwinm 32". @@ -2552,7 +2581,8 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { if (isRunOfOnes(Imm, MB, ME)) { SDValue Ops[] = { N->getOperand(0).getOperand(0), N->getOperand(0).getOperand(1), - getI32Imm(0), getI32Imm(MB),getI32Imm(ME) }; + getI32Imm(0, dl), getI32Imm(MB, dl), + getI32Imm(ME, dl) }; return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops); } } @@ -2593,7 +2623,8 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && isRotateAndMask(N, Imm, true, SH, MB, ME)) { SDValue Ops[] = { N->getOperand(0).getOperand(0), - getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; + getI32Imm(SH, dl), getI32Imm(MB, dl), + getI32Imm(ME, dl) }; return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); } @@ -2605,7 +2636,8 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) && isRotateAndMask(N, Imm, true, SH, MB, ME)) { SDValue Ops[] = { N->getOperand(0).getOperand(0), - getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) }; + getI32Imm(SH, dl), getI32Imm(MB, dl), + getI32Imm(ME, dl) }; return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops); } @@ -2625,11 +2657,12 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo; SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue, N->getOperand(0), - CurDAG->getTargetConstant(1, InVT)), 0); + CurDAG->getTargetConstant(1, dl, InVT)), + 0); SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32); SDValue SRIdxVal = CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ? - PPC::sub_eq : PPC::sub_gt, MVT::i32); + PPC::sub_eq : PPC::sub_gt, dl, MVT::i32); return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1, CR0Reg, SRIdxVal, @@ -2656,7 +2689,7 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { N->getValueType(0) == MVT::i32) { SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue, - N->getOperand(0), getI32Imm(~0U)); + N->getOperand(0), getI32Imm(~0U, dl)); return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(Tmp, 0), N->getOperand(0), SDValue(Tmp, 1)); @@ -2701,12 +2734,21 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { else if (N->getValueType(0) == MVT::i64) SelectCCOp = PPC::SELECT_CC_I8; else if (N->getValueType(0) == MVT::f32) - SelectCCOp = PPC::SELECT_CC_F4; + if (PPCSubTarget->hasP8Vector()) + SelectCCOp = PPC::SELECT_CC_VSSRC; + else + SelectCCOp = PPC::SELECT_CC_F4; else if (N->getValueType(0) == MVT::f64) if (PPCSubTarget->hasVSX()) SelectCCOp = PPC::SELECT_CC_VSFRC; else SelectCCOp = PPC::SELECT_CC_F8; + else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f64) + SelectCCOp = PPC::SELECT_CC_QFRC; + else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f32) + SelectCCOp = PPC::SELECT_CC_QSRC; + else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4i1) + SelectCCOp = PPC::SELECT_CC_QBRC; else if (N->getValueType(0) == MVT::v2f64 || N->getValueType(0) == MVT::v2i64) SelectCCOp = PPC::SELECT_CC_VSRC; @@ -2714,7 +2756,7 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { SelectCCOp = PPC::SELECT_CC_VRRC; SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3), - getI32Imm(BROpc) }; + getI32Imm(BROpc, dl) }; return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops); } case ISD::VSELECT: @@ -2748,7 +2790,8 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { DM[1] = 1 - tmp; } - SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), MVT::i32); + SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), dl, + MVT::i32); if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 && Op1.getOpcode() == ISD::SCALAR_TO_VECTOR && @@ -2787,7 +2830,7 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { // Op #4 is the Flag. // Prevent PPC::PRED_* from being selected into LI. SDValue Pred = - getI32Imm(cast(N->getOperand(1))->getZExtValue()); + getI32Imm(cast(N->getOperand(1))->getZExtValue(), dl); SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3), N->getOperand(0), N->getOperand(4) }; return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); @@ -2817,7 +2860,7 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { } SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl); - SDValue Ops[] = { getI32Imm(PCC), CondCode, + SDValue Ops[] = { getI32Imm(PCC, dl), CondCode, N->getOperand(4), N->getOperand(0) }; return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops); } @@ -2836,8 +2879,8 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { "Only supported for 64-bit ABI and 32-bit SVR4"); if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) { SDValue GA = N->getOperand(0); - return CurDAG->getMachineNode(PPC::LWZtoc, dl, MVT::i32, GA, - N->getOperand(1)); + return transferMemOperands(N, CurDAG->getMachineNode(PPC::LWZtoc, dl, + MVT::i32, GA, N->getOperand(1))); } // For medium and large code model, we generate two instructions as @@ -2857,12 +2900,12 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { SDValue GA = N->getOperand(0); SDValue TOCbase = N->getOperand(1); SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64, - TOCbase, GA); + TOCbase, GA); if (isa(GA) || isa(GA) || CModel == CodeModel::Large) - return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA, - SDValue(Tmp, 0)); + return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl, + MVT::i64, GA, SDValue(Tmp, 0))); if (GlobalAddressSDNode *G = dyn_cast(GA)) { const GlobalValue *GValue = G->getGlobal(); @@ -2870,8 +2913,8 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { (GValue->isDeclaration() || GValue->isWeakForLinker())) || GValue->isDeclaration() || GValue->hasCommonLinkage() || GValue->hasAvailableExternallyLinkage()) - return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA, - SDValue(Tmp, 0)); + return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl, + MVT::i64, GA, SDValue(Tmp, 0))); } return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64, @@ -2920,7 +2963,7 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { // Into: tmp = VSPLTIS[BHW] elt // VADDU[BHW]M tmp, tmp // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4 - SDValue EltVal = getI32Imm(Elt >> 1); + SDValue EltVal = getI32Imm(Elt >> 1, dl); SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); SDValue TmpVal = SDValue(Tmp, 0); return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal); @@ -2932,9 +2975,9 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { // Into: tmp1 = VSPLTIS[BHW] elt-16 // tmp2 = VSPLTIS[BHW] -16 // VSUBU[BHW]M tmp1, tmp2 - SDValue EltVal = getI32Imm(Elt - 16); + SDValue EltVal = getI32Imm(Elt - 16, dl); SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); - EltVal = getI32Imm(-16); + EltVal = getI32Imm(-16, dl); SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0), SDValue(Tmp2, 0)); @@ -2946,9 +2989,9 @@ SDNode *PPCDAGToDAGISel::Select(SDNode *N) { // Into: tmp1 = VSPLTIS[BHW] elt+16 // tmp2 = VSPLTIS[BHW] -16 // VADDU[BHW]M tmp1, tmp2 - SDValue EltVal = getI32Imm(Elt + 16); + SDValue EltVal = getI32Imm(Elt + 16, dl); SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); - EltVal = getI32Imm(-16); + EltVal = getI32Imm(-16, dl); SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal); return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0), SDValue(Tmp2, 0)); @@ -3157,7 +3200,8 @@ SDValue PPCDAGToDAGISel::combineToCMPB(SDNode *N) { bool NonTrivialMask = ((int64_t) Mask) != INT64_C(-1); if (NonTrivialMask && !Alt) { // Res = Mask & CMPB - Res = CurDAG->getNode(ISD::AND, dl, VT, Res, CurDAG->getConstant(Mask, VT)); + Res = CurDAG->getNode(ISD::AND, dl, VT, Res, + CurDAG->getConstant(Mask, dl, VT)); } else if (Alt) { // Res = (CMPB & Mask) | (~CMPB & Alt) // Which, as suggested here: @@ -3166,13 +3210,81 @@ SDValue PPCDAGToDAGISel::combineToCMPB(SDNode *N) { // Res = Alt ^ ((Alt ^ Mask) & CMPB) // useful because the (Alt ^ Mask) can be pre-computed. Res = CurDAG->getNode(ISD::AND, dl, VT, Res, - CurDAG->getConstant(Mask ^ Alt, VT)); - Res = CurDAG->getNode(ISD::XOR, dl, VT, Res, CurDAG->getConstant(Alt, VT)); + CurDAG->getConstant(Mask ^ Alt, dl, VT)); + Res = CurDAG->getNode(ISD::XOR, dl, VT, Res, + CurDAG->getConstant(Alt, dl, VT)); } return Res; } +// When CR bit registers are enabled, an extension of an i1 variable to a i32 +// or i64 value is lowered in terms of a SELECT_I[48] operation, and thus +// involves constant materialization of a 0 or a 1 or both. If the result of +// the extension is then operated upon by some operator that can be constant +// folded with a constant 0 or 1, and that constant can be materialized using +// only one instruction (like a zero or one), then we should fold in those +// operations with the select. +void PPCDAGToDAGISel::foldBoolExts(SDValue &Res, SDNode *&N) { + if (!PPCSubTarget->useCRBits()) + return; + + if (N->getOpcode() != ISD::ZERO_EXTEND && + N->getOpcode() != ISD::SIGN_EXTEND && + N->getOpcode() != ISD::ANY_EXTEND) + return; + + if (N->getOperand(0).getValueType() != MVT::i1) + return; + + if (!N->hasOneUse()) + return; + + SDLoc dl(N); + EVT VT = N->getValueType(0); + SDValue Cond = N->getOperand(0); + SDValue ConstTrue = + CurDAG->getConstant(N->getOpcode() == ISD::SIGN_EXTEND ? -1 : 1, dl, VT); + SDValue ConstFalse = CurDAG->getConstant(0, dl, VT); + + do { + SDNode *User = *N->use_begin(); + if (User->getNumOperands() != 2) + break; + + auto TryFold = [this, N, User, dl](SDValue Val) { + SDValue UserO0 = User->getOperand(0), UserO1 = User->getOperand(1); + SDValue O0 = UserO0.getNode() == N ? Val : UserO0; + SDValue O1 = UserO1.getNode() == N ? Val : UserO1; + + return CurDAG->FoldConstantArithmetic(User->getOpcode(), dl, + User->getValueType(0), + O0.getNode(), O1.getNode()); + }; + + SDValue TrueRes = TryFold(ConstTrue); + if (!TrueRes) + break; + SDValue FalseRes = TryFold(ConstFalse); + if (!FalseRes) + break; + + // For us to materialize these using one instruction, we must be able to + // represent them as signed 16-bit integers. + uint64_t True = cast(TrueRes)->getZExtValue(), + False = cast(FalseRes)->getZExtValue(); + if (!isInt<16>(True) || !isInt<16>(False)) + break; + + // We can replace User with a new SELECT node, and try again to see if we + // can fold the select with its user. + Res = CurDAG->getSelect(dl, User->getValueType(0), Cond, TrueRes, FalseRes); + N = User; + ConstTrue = TrueRes; + ConstFalse = FalseRes; + } while (N->hasOneUse()); +} + void PPCDAGToDAGISel::PreprocessISelDAG() { SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode()); ++Position; @@ -3191,6 +3303,9 @@ void PPCDAGToDAGISel::PreprocessISelDAG() { break; } + if (!Res) + foldBoolExts(Res, N); + if (Res) { DEBUG(dbgs() << "PPC DAG preprocessing replacing:\nOld: "); DEBUG(N->dump(CurDAG)); @@ -3332,8 +3447,12 @@ void PPCDAGToDAGISel::PeepholeCROps() { case PPC::SELECT_I8: case PPC::SELECT_F4: case PPC::SELECT_F8: + case PPC::SELECT_QFRC: + case PPC::SELECT_QSRC: + case PPC::SELECT_QBRC: case PPC::SELECT_VRRC: case PPC::SELECT_VSFRC: + case PPC::SELECT_VSSRC: case PPC::SELECT_VSRC: { SDValue Op = MachineNode->getOperand(0); if (Op.isMachineOpcode()) { @@ -3639,8 +3758,12 @@ void PPCDAGToDAGISel::PeepholeCROps() { case PPC::SELECT_I8: case PPC::SELECT_F4: case PPC::SELECT_F8: + case PPC::SELECT_QFRC: + case PPC::SELECT_QSRC: + case PPC::SELECT_QBRC: case PPC::SELECT_VRRC: case PPC::SELECT_VSFRC: + case PPC::SELECT_VSSRC: case PPC::SELECT_VSRC: if (Op1Set) ResNode = MachineNode->getOperand(1).getNode(); @@ -3729,6 +3852,19 @@ static bool PeepholePPC64ZExtGather(SDValue Op32, return true; } + // LHBRX and LWBRX always clear the higher-order bits. + if (Op32.getMachineOpcode() == PPC::LHBRX || + Op32.getMachineOpcode() == PPC::LWBRX) { + ToPromote.insert(Op32.getNode()); + return true; + } + + // CNTLZW always produces a 64-bit value in [0,32], and so is zero extended. + if (Op32.getMachineOpcode() == PPC::CNTLZW) { + ToPromote.insert(Op32.getNode()); + return true; + } + // Next, check for those instructions we can look through. // Assuming the mask does not wrap around, then the higher-order bits are @@ -3916,6 +4052,9 @@ void PPCDAGToDAGISel::PeepholePPC64ZExt() { case PPC::SRW: NewOpcode = PPC::SRW8; break; case PPC::LI: NewOpcode = PPC::LI8; break; case PPC::LIS: NewOpcode = PPC::LIS8; break; + case PPC::LHBRX: NewOpcode = PPC::LHBRX8; break; + case PPC::LWBRX: NewOpcode = PPC::LWBRX8; break; + case PPC::CNTLZW: NewOpcode = PPC::CNTLZW8; break; case PPC::RLWIMI: NewOpcode = PPC::RLWIMI8; break; case PPC::OR: NewOpcode = PPC::OR8; break; case PPC::SELECT_I4: NewOpcode = PPC::SELECT_I8; break;